ルールベース画像処理のススメ

fkubota

June 18, 2021

Tweet Share

Transcript

1. ϧʔϧϕʔεը૾ॲཧͷεεϝ fkubota DeepLearning ωΠςΟϒੈ୅΁޲͚ͯ

2. ࣗݾ঺հ 02 fkubota (Twitter, Kaggle) - όϯυϧΧʔυͷձࣾ(ΧϯϜ)ͰػցֶशΤϯδχΞ - Kaggle Expert

   - ԭೄग़਎(౦ژʹग़͖ͯͯ3೥൒) - ෺ཧֶՊग़਎(ڧ૬ؔిࢠܥɺ4.2KͷӷମϔϦ΢Ϝʹ࣓ੑମಥͬࠐΜͰͨ) - ϓϩάϥϛϯάྺ͸2೥൒͙Β͍ - झຯ - ૣى͖(4࣌൒ىচ) - Kaggle - ίʔώʔɺϏʔϧɺ΢ΟεΩʔ - ಡॻɺ෺ཧɺ఩ֶ

3. ͬͦ͘͞Ͱ͕͢ ͋Δݐ෺ͷϑϩΞը૾ ͜Εʹ ͜Μͳײ͡Ͱ੺͍఺Λଧ͍ͪͨ - ੨͍ྖҬʹೖͬͪΌͩΊ - ௨࿏͚ͩ Έͳ͞ΜͳΒͲ͏͠·͔͢ʁ 03

4. ൃදͷϞνϕʔγϣϯ ࣮͸͜ͷը૾ ͸Indoorίϯϖ ͷը૾Ͱ ੺͍఺ΛଧͭͨΊʹ؆୯ͳը૾ॲཧΛͯ͠ଧͪ·ͨ͠ νʔϜϝΠτʹղઆ͢Δͱ൓ԠΑ͔ͬͨͷͰ ϧʔϧϕʔεը૾ॲཧʹ͍ͭͯൃද͠Α͏ͱࢥ͍·ͨ͠ KaggleΛ͖͔͚ͬʹը૾ॲཧɾը૾ೝࣝΛ͸͡Ίͨਓ͸͍͖ͳΓ DeepLearning͔Βೖͬͨͱࢥ͍·͕͢ྺ࢙తʹ͸ϧʔϧϕʔε͸ͷ΄͏ ͕௕͘ɺະͩ׆༂͍ͯ͠Δٕज़Ͱ͢ɻ

   ֶΜͰଛ͸ͳ͍͸ͣʂʂ 04

5. ࠓ೔࿩͢͜ͱ σΟʔϓϥʔχϯάωΠςΟϒੈ୅ʹ޲͚ͯʂ 05 ը૾ͬͯͳʹʁ ը૾ॲཧೖ໳ Indoorͷ৔߹ ͦ΋ͦ΋ը૾ͬͯͳʹ͔Λ੔ ཧ͠·͢ɻ ϧʔϧϕʔεը૾ॲཧʹͪΐͬ ͱ͚ͩೖ໳ͯ͠Έ·͢ɻ

   opencvΛ࢖͍·͢ɻ https://github.com/fkubota/ bunseki_compe_LT_02/blob/main/ notebook/nb01_opencv.ipynb IndoorͰ΍ͬͨ͜ͱΛ΄Μͷগ ͠঺հ͠·͢ɻ

6. ը૾ͬͯԿʁ

7. ը૾Λ؆୯ʹ - pixel ͱ͍͏࠷খ୯ҐΛ΋ͭ - ࠲ඪx, y Λ΋ͭ - ً౓஋Λ΋ͭ

   - ௨ৗ͸8bit - 0~255ͷ஋ΛऔΔ(256ஈ֊) - 0͕ࠇ - 255͕ന y x നͬΆ͍ͷͰ 255ʹ͍ۙ஋ ࠇͳͷͰ0ʹ͍ۙ஋ փ৭ͳͷͰਅΜதͷ 150͙Β͍ʁ ΧϥʔͩͱRGBͷ3͕ͭͦΕͧΕ8bitͷ৘ใΛ࣋ͭ ྫ: - (133, 30, 88) ←10ਐ਺ - #56a53f ←16ਐ਺ 07

8. ը૾ॲཧೖ໳

9. ϧʔϧϕʔεͷ͍͍ͱ͜Ζ ڭࢣσʔλ͍Βͳ͍ ָ͍͠ʂ - ΍ͬͨΒΘ͔Δʂʂʂʂ - ͜Εҙ֎ͱݟམͱ͞Ε͕ͪ 09

10. ໰1 from fkubota shibutsu dataset Έͳ͞ΜͳΒͲ͏͠·͔͢ʁ (ڭࢣσʔλͳ͠Ͱ͢Αʁ) Ωϟοϓͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞ ༡ͼײ֮Ͱʂʂ 10

    https://github.com/fkubota/ bunseki_compe_LT_02/blob/main/ notebook/nb01_opencv.ipynb શ෦githubʹ͋ΔΑ

11. ղ๏ͷ1ྫΛ঺հ ·ͣ͸ఆ൪ͷ2஋ԽΛ͢ΔͨΊʹ άϨʔը૾ʹม׵͠·͢ʂ ֤pixelͷً౓஋(r, g, b) ͷ3࣍ݩ͕ 1࣍ݩʹͳΓ͍Ζ͍Ζѻ͍΍͘͢ͳΓ·͢ɻ 11

12. 2஋Խ എܠͱ෺ମΛ෼͚·͠ΐ͏ ً౓஋ͷώετάϥϜΛݟͯΈΔͱ - നͬΆ͍৭(255ʹ͍ۙ৭)͕ώετάϥ ϜͷϚδϣϦςΟʹͳͬͯΔ - ͜Ε͸എܠ͕ը૾ͷେ෦෼Λ઎ΊΔ͜ ͱʹىҼ͢Δ -

    threshold = 70 ͷલޙͰ2஋Խ͢Δʂ 1 0 Ωϟοϓ എܠ 12

13. 2஋Խ݁Ռ ϐϯΫͷ෦෼͕1 ͦΕҙ֎͕0 ͷ ΦϒδΣΫτ ͕Ͱ͖ͨ ͜ΕΛ “ྖҬ” ͱݴ͍·͢ 13

14. ͋ͱ͸៉ྷʹ͢Δ ࠓճ͸݀ΛຒΊ͚ͨͩ 14

15. ॏ৺Λܭࢉ ྖҬͷྠֲΛऔͬͯॏ৺Λܭࢉ Ͱ͖ͨʂʂ 15

16. ໰2 from fkubota shibutsu dataset Ωϟοϓͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞ ָ͘͠ͳ͖ͬͯ·ͨ͠ʁ 16

17. Կ͕೉͍͠ʁ Ωϟοϓͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞ ಉ͡໰୊ͳΜ͚ͩͲෳ਺෺ମ͕͋ΔͷͰ ೉қ౓͕௓Ͷ্͕Γ·͢Ͷɻ 17

18. ఆ൪ͷॲཧ άϨʔը૾ 2஋Խ 18

19. νϣοτ͖Ε͍ʹ ϞϧϑΥϩδʔม׵ - ࣌ؒͷ౎߹্ৄࡉ͸ল͖·͢ - ϊΠζআڈɺ݀ຒΊ͕Ͱ͖·͢ - ݪཧ͸؆୯ͳͷͰͥͻௐ΂ͯΈͯ ͍ͩ͘͞ ݀ຒΊ

    ΊͬͪΌ͖Ε͍ʹͳͬͨʂʂ 19

20. ϥϕϦϯά ࿈ଓͨ͠ΦϒδΣΫτຖʹ෼ׂ - ྡΓ߹ͬͨ1ΛಉҰͷϥϕϧͱͯ͠෼ׂ͢Δ - ϐϯΫͷྖҬ͕ෳ਺ͷྖҬʹ෼ׂ͞Εͨ - ෼ׂޙ໘ੵ͕খ͍͞΋ͷΛϊΠζͱͯ͠আ͍ͨ ໘ੵͷখ͍͞ྖҬ(ϊΠζ) 20

21. ͓ΘΓ͡Όͳ͍Αʂ ͜ͷத͔ΒΩϟοϓΛબͼ ग़͞ͳ͍ͱ͍͚·ͤΜ ಛ௃ྔΛ࢖͍·͢ ͜ͷ5ͭͷ෺ମΛൺֱͯ͠Ωϟοϓʹ͸Ͳ͏ ͍ͬͨಛ௃͕͋ΔͰ͠ΐ͏͔ʁ - ؙΈ͕͋Δɹ - ໘ੵ͕খ͍͞

    ͕ࢥ͍౰ͨΓ·͢ɻ ໘ੵ͕খ͍͞Λ࢖ͬͯΈ·͢ɻ ͭ·Γ ಛ௃ྔʹ໘ੵΛ࢖͏ͱ͍͏͜ͱʹͳΓ·͢ɻ 21

22. ྖҬͷ໘ੵΛܭࢉ ໘ੵ࠷খΛબͿ ಛ௃ྔ = ໘ੵ Ͱ͖ͨʂʂ 22

23. ໰3 Ͱ͖ͦʁ from fkubota shibutsu dataset ϖϯͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞ 23

24. ϥϕϦϯά·Ͱ͸Ұॹ ϖϯͷಛ௃Λߟ͑Δ fkubota͞Μ͸ɺʮࡉ௕͍ʯҙ֎ ࢥ͍͖ͭ·ͤΜͰͨ͠ɻ ࡉ௕͍͸͜͏දݱ͢Δ͜ͱʹ͠·͢ɻ 1. ྖҬΛ࠷খ֎઀ۣܗͰғ͏ 2. ʮ௕ลͷ௕͞ʯΛʮ୹ลͷ௕͞ʯͰׂΔ 3.

    ͜ΕΛʮࡉ௕͞౓ʯͱ͢Δ 24

25. ࡉ௕͍౓Λܭࢉ ಛ௃ྔ = ࡉ௕͍౓ ࠷େ஋ΛબͿ Ͱ͖ͨʂʂ 25

26. ໰4 ࠷ޙͩΑʂ from fkubota shibutsu dataset MINTIAͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞ 26

27. ಛ௃͸ʁ MINTIAͷಛ௃Λߟ͑Δ ࢛͍֯ͱ͍͏ಛ௃͕͋Γ·͕͢ɺ ଞͷ෺ମͱࠩผԽ͕೉͍͠Ͱ͢ɻ ܗͷ؍఺Ͱ͸೉ͦ͠͏ͳͷͰ ʮ৭ʯͰ߈ΊͯΈ·͠ΐ͏ɻ ʮ੨͞ʯΛදݱ͠·͢ɻ RGBΛHSVʹม׵͠ H(hue, ৭૬)ͷը૾Λ࢖͍·͢ɻ

    27

28. ৭૬ʁ ͜ͷล͕੨ ৭૬͸৭ͷ༷૬Λ 0~180·ͰͰมԽͤͨ͞΋ͷɻ ੨͸100~140͋ͨΓɻ 28

29. Hueը૾ hueը૾ ͜ͷลΓ 29

30. 2஋Խͯ͠ϞϧϑΥϩδʔ 2஋Խ ϞϧϑΥϩδʔ 30

31. ॏ৺Λܭࢉͯ͠ऴΘΓʂ Ͱ͖ͨʂʂ 31

32. Indoorͷ৔߹

33. ࠷ॳͷ࣭໰ ͋Δݐ෺ͷϑϩΞը૾ ͜Εʹ ͜Μͳײ͡Ͱ੺͍఺Λଧ͍ͪͨ - ੨͍ྖҬʹೖͬͪΌͩΊ - ௨࿏͚ͩ Έͳ͞ΜͳΒͲ͏͠·͔͢ʁ 33

34. ·ͣ͸௨࿏Λऔಘ ੨͍෦෼Λ2஋Խ ຒΊΔ ന͍෦෼Λ 2஋Խ ࿦ཧੵ 34

35. ׬੒ʂ & ੺͍఺ʑΛॻ͍ͨը૾Λ༻ҙͯ͠ ࿦ཧੵ Ͱ͖ͨʂʂ 35

36. Thanks :)